Decision Tree Analysis to Identify Factors that Impact Methylmercury Fraction in Wet Deposition

Transcription

1 Decision Tree Analysis to Identify Factors that Impact Methylmercury Fraction in Wet Deposition Dennis G. Jackson P.E. and Stephen P. Harris Ph.D. Savannah River National Laboratory 2016 Annual Meeting and Scientific Symposium of the National Atmospheric Deposition Program. October 31 November 4, 2016, La Fonda on the Plaza, Santa Fe, NM

2 NADP S MDN Methylmercury Dataset RELEASED FALL ,050 OBSERVATIONS 61 NADP-MDN SITES OCT 1996 FEB

3 Previous Assessments of the Methylmercury Dataset Wetherbee, Rhodes, Gay, Brunette, Prestbo, & Risch 2015 a) Poster at the International Conference of Mercury as a Global Pollutant Jeju, Korea b) Panel Presentation at the 2015 AGU-CGU Joint Meeting Montreal, Canada Mean concentration of ± ng/l Median concentration of ng/l Debris shifted the mean to ± ng/l and median of ng/l 0.5 to 4.0% of Total Hg is Methyl species Seasonal variation exists with maximum concentrations occurring in the summer (Wetherbee et al. 2015a) 3

4 Sources of Methylmercury in Precipitation Potential source(s) of methylmercury in wet deposition. Volatilization of monomethylmercury, Evasion and demethylation of dimethylmercury, Direct methylation of Hg 0 in the atmosphere. Gardfedlt et al. (2003): Acetic acid as an abiotic methylating agent in atmospheric and surface waters. Presence of chloride, oxalate, and sulfite that may limit methylation rate due to competition between acetate and Hg(II). Celo et al. (2006): Contributions of various methyl donors (Co, Sn, I) on abiotic methylation. Methylation is dependent upon ph, temperature, and complexing agents especially chloride. Q: What is the role of major cations and anions on the presence of methymercury? 4

5 Methymercury & NTN Sites 6,050 2,236 OBSERVATIONS MDN & NTN SITES 5

6 Methymercury & NTN Sites with matched events AVAILABLE MATCHED SAMPLES FL11 1X 4.27 ng/l IN20 16X 0.34 ng/l IN34 9X 0.44 ng/l MD08 1X 0.62 ng/l MS12 1X 0.11 ng/l 6,050 2,236 OBSERVATIONS MDN & NTN SITES 6

7 Exploratory Statistics - Decision Tree Analysis (Learning) Decision Tree Analysis recursively partitions data to create a tree of partitions. Groupings of X values are identified that best predict the Y value, All possible cuts are searched to optimize statistical parameters, Splits are done recursively forming a decision tree until an optimal fit is reached, Process chooses optimum splits from a large number of possible splits. Goal is to identify inorganic constituents that are associated with methylmercury in precipitation. Advantages of the Approach: Explore relationships without a good model, Process handles large problems easily, and Results are very interpretable. 7

8 Illustration of Method Matched NTN & MDN Weekly Samples % Methylmercury ALL DATA (n=28) 8

9 Illustration of Method First Split % Methylmercury [Mg] >= (n=14) ALL DATA (n=28) [Mg] < (n=13) 9

10 Illustration of Method Second Split % Methylmercury R 2 = RMSE = [Mg] < (n=9) [Mg] >= (n=6) [SO4] < (n=7) [SO4] >= (n=6) [Mg] >= (n=15) [Mg] < (n=13) ALL DATA (n=28) 10

11 Nuisances of the NADP Methylmercury Dataset Quality Code (n = 5,995) Reporting Limit {0.005, } (n = 5,995) Collection Period Days (n = 5,984) Number of Samples in Composites (n = 3,354; 56%) 11

12 Decision Tree for Volume Weighted Dataset (n=2,236) % Methylmercury ALL DATA (n =2,236) Mean = , σ = [NO 3 ] >= (n=244) Mean = , σ = Molar Concentrations (E-06) [SO4] >= (n=2,202) Mean = , σ = [K] < (n=1,179) Mean = , σ = [NO 3 ] < (n=935) Mean = , σ = [K] >= (n=1,023) Mean = , σ = [SO4] < (n=34) Mean = , σ =

13 Decision Tree for Volume Weighted Dataset (n=2,236) % Methylmercury ALL DATA (n =2,236) Mean = , σ = [NO 3 ] >= (n=244) Mean = , σ = [SO4] >= (n=2,202) Mean = , σ = [K] < (n=1,179) Mean = , σ = [NO 3 ] < (n=935) Mean = , σ = [K] >= (n=1,023) Mean = , σ = [SO4] < (n=34) Mean = , σ = Molar Concentrations (E-06) R 2 = 0.083; RMSE =

14 Summary & Conclusions NTN Constituents associated with Methylmercury are - Inconclusive Advantages of the Approach Leverage observations from multiple NADP networks to gain insight on variables that may influence methylmercury in precipitation. Improvement Opportunities Heavily censored Methyl data appropriate analysis techniques? Consider Site Location, Season, and Debris as components, Compositing masks details associated with weekly extreme events, Use of total mercury vs. reactive mercury (Hammerschmidt et al., 2007) 14

15 Questions & Acknowledgements Advice, counsel, and commiserating from Greg Wetherbee, David Gay, and Mark Rhodes Select Illustrations Sir John Tenniel 15